CLIP

CLIP learns a shared representation for pictures and phrases by training an image encoder and a text encoder to agree on which caption matches which image inside a large batch. After training, any image and any piece of text can be embedded into a common vector space, and simple cosine similarity decides how well they match. This makes it easy to build zero shot classifiers by writing label prompts, to search images with natural language, and to rerank or filter results from generative or retrieval systems. The approach is robust across many domains because the supervision comes from broad web data rather than a single labeled dataset, and it works without fine tuning for many tasks. In practice, teams use CLIP for image search, dataset curation, safety and content tagging, grounding for multimodal assistants, and as a scoring model that keeps outputs aligned with user intent.